🧠 Think MCP v4.6.2

MCP Server for structured sequential thinking. Helps LLMs break down complex problems into manageable steps with branching, revisions, cross-session learning, and proactive nudges.

📝 Base Prompt: pastebin.com/EJcJ3fJF

✨ Features

• 5 Streamlined Tools — think, think_batch, think_done, think_recall, think_reset
• Burst Thinking — Submit up to 30 thoughts in one call
• Cross-Session Learning — Insights from past sessions via think_recall(scope:insights)
• Proactive Nudges — Short prompts for self-reflection when patterns detected
• Branching & Revisions — Explore alternatives, fix mistakes
• Dead Ends Tracking — Remember rejected paths
• Fuzzy Recall — Search through thought history with Fuse.js
• Session Persistence — Auto-save/restore with 24h TTL

🚀 Quick Start

npx @gofman3/think-mcp

MCP Configuration

{
  "mcpServers": {
    "think": {
      "command": "npx",
      "args": ["-y", "@gofman3/think-mcp"]
    }
  }
}

🛠️ Tools

think

Add a single thought to the reasoning chain.

{
  thought: string,
  thoughtNumber: number,
  totalThoughts: number,
  nextThoughtNeeded: boolean,
  confidence?: number,        // 1-10
  goal?: string,              // Set in first thought
  subSteps?: string[],        // Micro-plan (max 5)
  alternatives?: string[],    // Quick comparison (max 5)
  quickExtension?: {
    type: 'critique' | 'elaboration' | 'correction' | 'innovation' | 'optimization' | 'polish',
    content: string,
    impact?: 'low' | 'medium' | 'high' | 'blocker'
  },
  isRevision?: boolean,
  revisesThought?: number,
  branchFromThought?: number,
  branchId?: string
}

think_batch

Submit complete reasoning chain in one call (1-30 thoughts).

{
  goal: string,               // Min 10 chars
  thoughts: [{
    thoughtNumber: number,
    thought: string,          // Min 50 chars
    confidence?: number,
    subSteps?: string[],
    alternatives?: string[],
    extensions?: [{ type, content, impact }]
  }],
  consolidation?: {
    winningPath: number[],
    summary: string,
    verdict: 'ready' | 'needs_more_work'
  }
}

think_done

Finish session with verification and optional export.

{
  winningPath: number[],
  summary: string,
  verdict: 'ready' | 'needs_more_work',
  exportReport?: 'markdown' | 'json',
  includeMermaid?: boolean
}

think_recall

Search current session or past insights.

{
  query: string,
  scope?: 'session' | 'insights',  // Default: session
  searchIn?: 'thoughts' | 'extensions' | 'alternatives' | 'all',
  limit?: number,
  threshold?: number
}

think_reset

Clear session and start fresh.

💡 Nudge System (v4.6)

When no warnings are present, the server returns short prompts based on detected patterns:

| Pattern | Nudge | |---------|-------| | confidence < 5 | "Low confidence. Validate assumptions?" | | 3+ thoughts without alternatives | "No alternatives explored. Tunnel vision?" | | Complex goal without subSteps | "Complex goal, no breakdown. Decompose?" | | Unresolved blocker | "Blocker unresolved. Address before continuing?" |

Nudges appear only when there's no other systemAdvice — avoiding noise.

📊 Complexity Budget

| Task | Thoughts | Tool | |------|----------|------| | Simple | 0-2 | Skip or think | | Medium | 3-7 | think step-by-step | | Complex | 8-30 | think_batch |

🔄 Changelog

v4.6.0

• New: NudgeService — proactive micro-prompts for self-reflection
• New: Nudge field in response (💡 icon)
• Improved: Nudges only appear when no other warnings present

v4.5.0

• Breaking: Renamed tools (sequentialthinking → think, etc.)
• New: Cross-session insights via think_recall(scope:insights)
• Improved: Compact output, lazy tree generation

v4.1.0

• New: submit_thinking_session (Burst Thinking)
• New: Atomic validation (sequence, stagnation, entropy)

v3.4.0

• New: Fuzzy recall with Fuse.js
• New: Dead ends tracking
• New: Session TTL (24h)

📄 License

MIT

🎯 How to Use (for humans)

Simple task — DON'T USE TOOLS

You: "Add console.log for debugging"
AI: *just does it, no thinking needed*

Medium/Complex task — FULL WORKFLOW

Example prompt:

Working directory: C:/Projects/my-backend

1. Study folders src/auth/ and src/utils/
2. use think_batch — analyze architecture and edge cases
3. use task_spec — create specification with tasks

Goal: "Build robust password validation backend logic"

Requirements:
- Minimum 8 characters
- At least 1 digit, 1 special character
- Check against leaked passwords (haveibeenpwned API)
- Rate limiting on validation

What happens:

1. AI studies specified folders
2. think_batch — thinks through edge cases (what if API is down? what about unicode passwords?)
3. task_spec — creates spec with tasks in .gofman3/specs/password-validation/
4. You'll see files:
   • requirements.md — what we're building
   • design.md — how we're building
   • tasks.md — progress (auto-updates!)

Track progress

Show task_board

AI shows Kanban board — what's done, in progress, blocked.

If AI is stuck

use think_recall query: "password validation" scope: insights

Searches past sessions for similar solutions.